⚡ Apex-1-flash
Fast. Sharp. Thinks Before It Speaks.
A chain-of-thought reasoning model by OrbitAI
Built by a 13-year-old developer from Slovakia — because curiosity has no age limit.
🔍 Overview
Apex-1-flash is a supervised fine-tune of
Qwen/qwen3-4b-thinking-2507, purpose-built to deliver
sharp, structured reasoning with efficient chain-of-thought capabilities at the 4B parameter scale.
Trained on the
Open-CoT-Reasoning-Mini dataset, apex-1-flash is designed to think through problems step by step — making it well-suited for logical reasoning, multi-step problem solving, and coherent explanations — while staying lean enough to run on consumer hardware.
This model was created by Matias Mikle (age 13, Slovakia 🇸🇰) alongside the OrbitAI team.
📋 Model Details
🧠 What Makes apex-1-flash Different
The name says it all — Apex for reaching the top, flash for speed and precision.
The flash philosophy shapes how the model was built:
- ⚡ Fast — At only ~4B parameters, it's lightweight enough to run on a single consumer GPU without sacrificing reasoning depth
- 🎯 Sharp — Fine-tuned specifically on structured chain-of-thought data, it breaks down problems cleanly before producing answers
- 💡 Thoughtful — Inherits the built-in thinking architecture from Qwen3, extended through CoT fine-tuning for more reliable step-by-step logic
Best suited for
- Logical and mathematical reasoning
- Step-by-step problem decomposition
- Structured explanation generation
- Research and educational tasks
- Multi-step Q&A
🚀 Quickstart
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4model_id = "OrbitAIEU/apex-1-flash"
5
6tokenizer = AutoTokenizer.from_pretrained(model_id)
7model = AutoModelForCausalLM.from_pretrained(
8 model_id,
9 torch_dtype=torch.bfloat16,
10 device_map="auto"
11)
12
13messages = [
14 {
15 "role": "user",
16 "content": "Explain step by step how to solve: 3x + 7 = 22"
17 }
18]
19
20text = tokenizer.apply_chat_template(
21 messages,
22 tokenize=False,
23 add_generation_prompt=True
24)
25
26inputs = tokenizer([text], return_tensors="pt").to(model.device)
27
28with torch.no_grad():
29 outputs = model.generate(
30 **inputs,
31 max_new_tokens=512,
32 temperature=0.7,
33 do_sample=True
34 )
35
36response = tokenizer.decode(
37 outputs[0][inputs.input_ids.shape[-1]:],
38 skip_special_tokens=True
39)
40print(response)
💾 Hardware Requirements
| Precision | Min. VRAM | Recommended For |
|---|
| Full precision (fp32) | ~16 GB | Not recommended |
| Half precision (bf16/fp16) | ~8 GB | RTX 3070 / RTX 4060 Ti and above |
| 4-bit quantized (GGUF/GPTQ) | ~3–4 GB | RTX 3060 / consumer-grade GPUs |
apex-1-flash is intentionally built at the 4B scale so it can run on everyday hardware — no enterprise cluster required.
🏋️ Training
The model was fine-tuned using Supervised Fine-Tuning (SFT) on top of the Qwen3-4B thinking checkpoint.
The Open-CoT-Reasoning-Mini dataset provides carefully structured reasoning traces and chain-of-thought examples, enabling the model to build stronger habits around multi-step logical inference.
⚠️ Limitations
- No safety alignment — Apex-1-flash has not undergone RLHF or safety tuning. It is not recommended for production use without additional safety layers.
- Domain scope — Performance is optimized for reasoning-heavy tasks; general-purpose capabilities are inherited from the base model.
- Inherited biases — The model may carry biases and limitations present in the Qwen3-4B base model.
- Benchmarks pending — Formal benchmark evaluations are currently in progress and will be published in a future update.
👤 About the Creator
Matias Mikle
Age: 13 · Country: Slovakia 🇸🇰
Independent developer, AI researcher, and founder of OrbitAI.
Matias started building AI projects from scratch, exploring fine-tuning, language model architecture, and full-stack development — proving that great work can come from anywhere, at any age.
|
"You don't need a Phd to train an AI model, you just need intelligence and GPU ofc."
🛰️ About OrbitAI
OrbitAI is an independent AI development team focused on building open, efficient, and accessible language models.
The team believes that AI research should not be limited to large corporations and well-funded labs. By working in the open — releasing models, sharing experiments, and collaborating with the community — OrbitAI aims to make frontier-style AI work accessible to anyone willing to put in the effort.
apex-1-flash is OrbitAI's first public model release.
📄 License
This model is released under the
Apache License 2.0, in accordance with the license of the base model
Qwen/qwen3-4b-thinking-2507.
| Permission | Allowed |
|---|
| Commercial use | ✅ Yes |
| Modification & distribution | ✅ Yes |
| Further fine-tuning | ✅ Yes |
| Research & academic use | ✅ Yes |
See the full
Apache 2.0 License for complete terms.
🙏 Acknowledgements
- Qwen Team @ Alibaba Cloud — for releasing the powerful Qwen3 model family under an open license
- Raymond-dev-546730 — for creating and sharing the Open-CoT-Reasoning-Mini dataset
- The open-source AI community — for making all of this possible
Apex-1-flash · Made with ❤️ by Matias Mikle & OrbitAI · Slovakia 🇸🇰
If this project inspired you — download it, fork it, and build something even better.
